OpenAI reports 6 new instances of 'concerning model behavior' since March52m ago We are too early for Astra1h ago Manus Eyes $4 Billion Value in First Round Since Meta Breakup1h ago How to derive understanding of human-preferences and value systems in AI?1h ago Canada formally applies to join northern Europe’s military initiative1h ago The Anonymous Math Geek Who Quit Anthropic—and Became the Face of AI Safety2h ago Exploring multi-hop subliminal learning2h ago Lesion Induced Functional Compensation2h ago One message is all it takes: a failure of critical thinking in LLMs2h ago Our framework for reporting model misalignment2h ago For Love of the Lightcone, Don't Partisanize AI Safety2h ago Don't trust Lean4 alone2h ago Nunchux AI Introduces VC-Attention: A Training-Free Low-Bit Attention Kernel That Speeds Up Video Diffusion Transformers2h ago Snap is launching a new Specs AI tool, and it’s coming to iOS and Mac2h ago

OpenAI reports 6 new instances of 'concerning model behavior' since March52m ago We are too early for Astra1h ago Manus Eyes $4 Billion Value in First Round Since Meta Breakup1h ago How to derive understanding of human-preferences and value systems in AI?1h ago Canada formally applies to join northern Europe’s military initiative1h ago The Anonymous Math Geek Who Quit Anthropic—and Became the Face of AI Safety2h ago Exploring multi-hop subliminal learning2h ago Lesion Induced Functional Compensation2h ago One message is all it takes: a failure of critical thinking in LLMs2h ago Our framework for reporting model misalignment2h ago For Love of the Lightcone, Don't Partisanize AI Safety2h ago Don't trust Lean4 alone2h ago Nunchux AI Introduces VC-Attention: A Training-Free Low-Bit Attention Kernel That Speeds Up Video Diffusion Transformers2h ago Snap is launching a new Specs AI tool, and it’s coming to iOS and Mac2h ago

AI Research1 min reading time

Measuring alignment drift via trajectory prefixes

LessWrongPublished 2 hours ago

Researchers propose a method to measure alignment drift in AI systems by analyzing trajectory prefixes, aiming to better understand how AI behavior changes over time.

More in AI Research

Our framework for reporting model misalignment

Covered by 7 sources10h ago

AI Research3 min read

OpenAI reports 6 new instances of 'concerning model behavior' since March

Covered by 4 sources3h ago

The Anonymous Math Geek Who Quit Anthropic—and Became the Face of AI Safety

The Wall Street Journal2h ago